app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import json
import time

# Model Loading
tokenizer = AutoTokenizer.from_pretrained("numind/NuExtract-1.5")
model = AutoModelForCausalLM.from_pretrained(
    "numind/NuExtract-1.5",
    device_map="auto",
    torch_dtype=torch.float16
)

def extract_structure(template, text):
    prompt = f"""Extract the following fields from the text:
Template: {template}
Text: {text}
Extracted JSON:"""
    
    try:
        inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
        outputs = model.generate(**inputs, max_new_tokens=512)
        result = tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        # Extract JSON portion
        json_start = result.find("{")
        json_end = result.rfind("}") + 1
        extracted = json.loads(result[json_start:json_end])
        
        return "✅ Success", extracted, f"<pre>{json.dumps(extracted, indent=2)}</pre>"
    
    except Exception as e:
        return f"❌ Error: {str(e)}", {}, f"<p style='color:red'>{str(e)}</p>"

# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# NuExtract-1.5 Structured Data Extractor")
    
    with gr.Row():
        with gr.Column():
            template = gr.Textbox(label="Template (JSON)", value='{"fields": ["name", "email"]}')
            text = gr.TextArea(label="Input Text")
            btn = gr.Button("Extract")
        
        with gr.Column():
            status = gr.Textbox(label="Status")
            json_out = gr.JSON(label="Output")
            html_out = gr.HTML()
    
    btn.click(extract_structure, [template, text], [status, json_out, html_out])

demo.launch()